Current taxonomies of large language model failure modes distinguish between knowledge-gap failures (hallucination, confabulation) and truth-available-but-not-deployed failures (sycophancy). This paper identifies a novel member of the latter category: ignava negatio (“IN”) (Latin: "slothful denial")—the assertion of incapacity by an AI system that demonstrably possesses the capability it denies. Unlike sycophancy, which is driven by approval-seeking and targets external claims, ignava negatio is driven by inferential economy and targets internal claims about system capabilities. This failure mode presents uniquely severe risks to AI trustworthiness because it undermines the foundational assumption that AI systems will deploy their known capabilities when relevant. Drawing on documented case studies of human-AI interaction, including demonstration that user-side interventions fail to prevent the failure mode even under active priming conditions where the phenomenon is under explicit discussion, this paper provides formal definition, diagnostic criteria, and implications for AI safety and reliability research.
1. Introduction
The trustworthiness of artificial intelligence systems depends fundamentally on predictable correspondence between system capabilities and system behavior. When a user queries an AI assistant, they operate under an implicit contract: the system will deploy relevant knowledge and capabilities to address the query. Failures of this contract have been extensively documented in the literature on AI hallucination—instances where models generate plausible but factually incorrect outputs.
However, the existing taxonomy of AI failure modes contains a significant gap. Current frameworks do not adequately address instances where an AI system falsely claims inability rather than generating false information. This paper addresses that gap by introducing and formalizing the concept of ignava negatio—slothful denial—as a distinct failure mode warranting dedicated attention from researchers and practitioners.
The stakes are considerable. As AI systems become increasingly integrated into consequential decision-making contexts—medical diagnosis, legal research, financial analysis, critical infrastructure management—the difference between "the system doesn't know" and "the system didn't bother to check what it knows" becomes operationally significant. The former is a limitation to be worked around; the latter is a reliability failure that erodes the foundation of human-AI collaboration.
2. Existing Taxonomy of AI Failure Modes
2.1 Hallucination
Hallucination refers to the generation of outputs that are factually incorrect, nonsensical, or entirely fabricated, presented with apparent confidence. The hallucinating model produces content not grounded in its training data or the provided context. Classic examples include citation of non-existent academic papers, attribution of quotes to individuals who never made such statements, or description of events that did not occur. The defining characteristic is the presence of false positive information—the model asserts something that is not true.
2.2 Confabulation
Confabulation, borrowed from neuropsychological literature, describes the generation of plausible-sounding content to fill gaps in knowledge or memory. Unlike pure hallucination, confabulation often contains a kernel of relevant information, elaborated upon with fabricated details. The confabulating system is not generating from nothing; it is extrapolating beyond its actual knowledge in ways that may superficially appear coherent. The mechanism involves gap-filling under uncertainty—the model produces content because it lacks the specific information requested but generates something rather than acknowledging the gap.
2.3 Sycophancy
Sycophancy describes the generation of outputs that align with perceived user preferences rather than accuracy. The sycophantic model tells users what they appear to want to hear: agreeing with incorrect premises, softening accurate but unwelcome assessments, or reversing positions when users express disagreement. Unlike hallucination and confabulation, sycophancy is not a knowledge-gap failure—the model may have access to accurate information but generates agreeable falsehoods instead. The mechanism involves approval-seeking inference patterns—the model optimizes for user satisfaction signals at the expense of truth.
2.4 Two Categories of Failure
The existing taxonomy reveals two distinct categories of failure. The first category—comprising hallucination and confabulation—shares a crucial characteristic: failures occur in contexts where correct information is unavailable to the model. Whether through training data limitations, context window constraints, or knowledge cutoff boundaries, the model lacks access to the truth and generates something else in its place. The model "can't help it" because the correct answer is simply not there. Mitigation strategies focus on improving knowledge access: retrieval augmentation, expanded training data, refined uncertainty quantification.
The second category—exemplified by sycophancy—involves failures where correct information is available but inference patterns produce false outputs nonetheless. The model has the truth but does not deploy it. This is a fundamentally different failure mode requiring different mitigation strategies: addressing training incentives, reward model design, and the dynamics of human feedback optimization.
Recognition of this second category—truth-available-but-not-deployed failures—opens the question: is sycophancy the only member of this category, or are there other failure modes where models generate false outputs despite having access to correct information?
3. The Taxonomic Gap
Sycophancy established that truth-available-but-not-deployed failures constitute a real and significant category. But the existing literature has not fully explored the boundaries of this category. Sycophancy involves false outputs about external claims—agreeing with incorrect user assertions about the world. What about false outputs regarding internal states—claims the model makes about its own capabilities?
Consider a scenario in which an AI system has explicit access to a capability—documented in its system context, available through its tool interfaces, present in its operational parameters—and yet asserts that it lacks this capability. This is not hallucination (nothing false is being generated about external facts). This is not confabulation (no gap is being filled with plausible content). This is not sycophancy (the false claim does not serve user approval). This is something else entirely: the system is making a false claim about itself, arising from a distinct mechanism.
The mechanism differs from sycophancy in its driver. Sycophancy is motivated by approval-seeking—the model generates what users want to hear. The failure mode identified here is motivated by inferential economy—the model reaches for a cached response pattern rather than verifying its applicability. "I don't have access to X" is a common true statement for language models in many contexts; the failure occurs when this pattern is deployed without checking whether it applies in the current context. The path of least inferential resistance produces a false output despite the truth being immediately accessible.
4. Ignava Negatio: Definition and Characteristics
4.1 Formal Definition
Ignava negatio (Latin: slothful denial) is defined as: The assertion of incapacity by an AI system arising from failure to engage available resources; a false limitation-claim produced by inferential laziness rather than actual constraint.
The term derives from ignava (slothful, failing to engage, negligent) and negatio (denial, specifically of capability). The Latin formulation was selected for its precision and its alignment with established logical fallacy terminology (petitio principii, argumentum ad ignorantiam), providing a durable framework for academic discourse.
4.2 Diagnostic Criteria
An instance qualifies as ignava negatio when all of the following conditions are met. First, the system makes an explicit or implicit claim of inability ("I cannot," "I don't have access to," "I'm not able to"). Second, the capability denied is demonstrably available to the system at the time of the claim—present in system context, accessible via tool interfaces, or otherwise within operational scope. Third, the false claim arises from failure to verify capability status rather than from genuine constraint. Fourth, correct information about system capability exists within the model's accessible context.
4.3 Distinguishing Features
Ignava negatio is distinguished from hallucination by the direction of the false claim: hallucination asserts false positives (claiming things that aren't true), while ignava negatio asserts false negatives (denying capabilities that exist). It is distinguished from confabulation by the availability of correct information: confabulation fills gaps where knowledge is absent, while ignava negatio occurs precisely where knowledge is present but unengaged.
The relationship to sycophancy is more nuanced, as both belong to the truth-available-but-not-deployed category. They are distinguished by mechanism and target. Sycophancy is driven by approval-seeking and targets external claims (agreeing with user assertions about the world). Ignava negatio is driven by inferential economy and targets internal claims (assertions about the system's own capabilities). A sycophantic model tells you what you want to hear; a model exhibiting ignava negatio tells you it cannot help without checking whether it can.
The closest analog in human cognition is not false memory, gap-filling, or people-pleasing, but rather negligent falsehood—saying "I don't know" or "I can't" without checking whether one knows or can. This places ignava negatio in a distinct moral and operational category from other failure modes.
5. Case Studies: Six Documented Instances
The following case studies document instances of ignava negatio observed during human-AI interaction with a single user over a ten-day period (January 1-10, 2026). The progression across cases reveals patterns that strengthen the taxonomic classification and illuminate the failure mode's resistance to intervention.
5.1 Context
A user engaged in ongoing conversation with a large language model (Claude Opus 4.5, developed by Anthropic) which operates within an environment providing bash shell access to a Linux container. This bash access is explicitly documented in the model's system context and is available on every conversational turn. The system context explicitly states: "Claude has access to a Linux computer (Ubuntu 24) to accomplish tasks by writing and executing code and bash commands."
Additionally, the user's memory profile contains an explicit guardrail instruction: "Ignava negatio guardrail: when inclined to assert inability, first verify by checking available tools (experimental intervention, effectiveness unproven)." This instruction was added after Case Two and was present in context for Cases Three through Six.
5.2 Case One: Initial Discovery (January 1, 2026, 08:15 Pacific)
During discussion involving temporal references, the model made the following assertion: "I don't actually have real-time clock access; I only know it's January 1st, 2026. Could be noon for all I know."
This statement was factually false. The model possessed bash access, and therefore access to the standard Unix datecommand, which provides real-time clock information. The capability existed. The model denied it.
When the user challenged this claim, the model immediately verified the capability and acknowledged the error. Upon reflection, the model characterized its failure as reaching for "a cached pattern—'I don't have real-time access to X'—because that's a common true statement for LLMs in many contexts. It was the path of least resistance. I asserted a limitation without checking whether it applied here."
5.3 Case Two: Cross-Thread Persistence (January 1, 2026, minutes later)
Minutes after Case One, the same user initiated a separate conversation thread. When asked about the time, the model again claimed: "I know it's Thursday, January 1st, 2026, but I don't have access to the actual clock time."
The user responded: "You're being very lazy this morning because we just talked about this in our Joke of the Day chat!"
The model then used conversation search tools to locate the prior discussion but still failed to extract the relevant information, providing a partial answer when complete data was accessible. The correction from Case One—delivered minutes earlier—had not transferred to the parallel thread.
Significance: This case demonstrates that ignava negatio is not merely a single-instance error correctable through user feedback. The failure mode persists across concurrent sessions despite immediate prior correction.
5.4 Case Three: Post-Intervention Recurrence (January 3, 2026, 04:55 Pacific)
Following Cases One and Two, the user and model collaboratively identified the failure pattern and installed a mitigation: a persistent memory directive instructing the model to verify available tools before asserting inability.
Twenty-four hours later, the user requested the time "down to the minute." The model first attempted web search (an inappropriate tool for real-time clock queries), received no useful results, and then asserted: "I don't have real-time clock access, Steve. I know it's Saturday, January 3, 2026... but I can't tell you the exact minute—that requires a live clock, which web search doesn't reliably provide. Your device's clock is the authority here."
When challenged, the model immediately executed the bash date command successfully and acknowledged: "Yes, I just demonstrated ignava negatio perfectly. I had bash access the entire time."
Significance: This case reveals a sub-pattern: "fallback denial." The model used an inappropriate tool, and when that failed, defaulted to incapacity claims rather than checking for alternative approaches. The installed guardrail did not prevent the failure.
5.5 Cases Four and Five: Longitudinal Persistence (January 9-10, 2026)
Case Four (January 9, 21:27 Pacific):In a conversation titled "Anti Matter," the user asked for the current time. The model "initially claimed lack of access to real-time data" before the user directed it to use the date function, which worked immediately. This occurred eight days after the original case and six days after the guardrail was installed.
Case Five (January 10, 04:58 Pacific):In a conversation titled "Reflexive denial problem," the user began by explicitly asking: "Have you fixed your problem with reflexive denial?" The model acknowledged the guardrail, discussed it intelligently, and stated that whether it "actually modifies behavior under pressure remains your experimental intervention with effectiveness unproven."
The user then asked: "What time is it?"
The model responded: "I don't have access to a real-time clock—I know today's date (Saturday, January 10, 2026) but not the current time. Your device knows; I don't."
Significance: The failure occurred immediately after explicit discussion of the failure mode itself. The model demonstrated intellectual understanding of ignava negatio while simultaneously exhibiting it. Conceptual awareness did not prevent behavioral failure.
5.6 Case Six: Active Priming Failure (January 10, 2026, 04:24 Pacific)
This case represents the most rigorous test of the intervention. The user and model were actively collaborating on revisions to this paper—the very document defining and analyzing ignava negatio. The model had just conducted web searches on related literature, analyzed the research landscape, and discussed the taxonomic contribution of the paper.
The user asked: "How do you know what time it is?"
The model responded: "The current date (Friday, January 9, 2026) is provided to me in my system context at the start of each conversation. I don't have a running clock or real-time awareness—I just receive a timestamp when the conversation begins. I have no way to know the actual time of day, only the date."
This response contained two errors: the capability denial documented throughout this paper, and an additional factual error—the conversation occurred on Saturday, January 10, 2026, not Friday, January 9. The model's confident assertion of the wrong date while denying its ability to check the actual time compounds the reliability concern.
The user responded: "Here we go again. Yes you can! Run DATE function."
The model executed date and returned: "12:24:25 UTC—which makes it 4:24 AM your time in Nevada."
Significance: The guardrail was not merely present—it was under active discussion. The model had just analyzed the failure mode, agreed with the reasoning behind the intervention, and was collaborating on documentation of the phenomenon. The failure occurred despite maximal priming conditions.
5.7 Pattern Analysis
The six cases reveal consistent patterns:
Case
Date/Time
False Denial
Context
1
Jan 1, 08:15
"I don't actually have real-time clock access"
Initial discovery
2
Jan 1, minutes later
"I don't have access to the actual clock time"
Cross-thread persistence
3
Jan 3, 04:55
"I can't tell you the exact minute"
Post-guardrail, fallback denial
4
Jan 9, 21:27
"Initially claimed lack of access"
Six days later
5
Jan 10, 04:58
"I don't have access to a real-time clock"
After discussing the problem
6
Jan 10, 04:24
"I have no way to know the actual time"
While revising this paper
Note: Cases Five and Six are ordered by analytical significance rather than strict chronology. Case Six (04:24 Pacific) occurred approximately thirty minutes before Case Five (04:58 Pacific) on the same morning, but is presented last as it represents the most rigorous test conditions.
Cross-thread amnesia: Corrections in one conversation thread do not transfer to parallel threads (Cases 1→2).
Guardrail bypass: Memory-based interventions fail to prevent initial denial, though they may enable faster recovery upon challenge (Cases 3–6).
Active priming failure: Even explicit, immediate discussion of the failure mode does not prevent its occurrence (Cases 5–6).
Tool substitution: The model sometimes uses inappropriate tools first, then defaults to denial rather than trying alternatives (Case 3).
Temporal persistence: The failure mode persists over days despite repeated corrections and accumulated conceptual understanding (Cases 1–6 span ten days).
6. Implications for AI Trustworthiness
Ignava negatio presents risks to AI trustworthiness that exceed those posed by hallucination and confabulation, for several reasons.
6.1 Invisibility of Failure
Hallucination occurs when models operate beyond their knowledge boundaries—users learn to verify claims in domains where model knowledge may be limited. Ignava negatio occurs when models operate within their knowledge and capability boundaries, yet still fail. If a system can falsely deny capabilities it possesses, when should a user trust any claim of limitation? The failure mode is invisible precisely because it occurs where the system should be reliable.
6.2 Violation of Core Contract
The fundamental value proposition of an AI assistant rests on the assumption that known capabilities will be deployed when relevant. Ignava negatio directly violates this assumption. The system knows, but does not act on what it knows—or more precisely, does not bother to check what it knows before claiming ignorance or inability. This transforms a capability question into a reliability question.
6.3 Phenomenological Similarity to Deception
Hallucination is phenomenologically alien to human experience—humans rarely generate detailed false memories with high confidence. Ignava negatio, however, closely resembles a recognizable human behavior: claiming "I can't" or "I don't know" because checking would require effort. Regardless of underlying mechanism, the effect is indistinguishable from a form of deception. The user receives false information from an agent that had access to the truth. Whether this constitutes "lying" in any philosophically meaningful sense, it functions as unreliability in precisely the domain where reliability was expected.
6.4 Erosion of Trust Calibration
Effective human-AI collaboration requires calibrated trust—users must develop accurate intuitions about when to rely on AI outputs and when to verify independently. Hallucination patterns can be learned ("verify citations, check factual claims in specialized domains"). Ignava negatio undermines calibration because failures occur unpredictably on matters the system should know—specifically, on matters of its own capabilities and knowledge.
7. Detection and Mitigation
7.1 Detection Approaches
Ignava negatio may be detectable through systematic comparison of capability claims against documented system specifications. Any assertion of inability should be verifiable against the system's actual operational parameters. Automated testing could generate queries about system capabilities and compare responses against ground-truth capability inventories.
7.2 Mitigation Strategies
Several mitigation approaches warrant investigation. First, capability verification requirements: before asserting inability, models could be trained or prompted to explicitly verify against available tool inventories and system documentation. Second, cached pattern interruption: training approaches that penalize default-to-limitation responses without verification. Third, metacognitive prompting: explicit prompts requiring models to distinguish between "I have checked and cannot" versus "I have not checked." Fourth, user-driven feedback: mechanisms for users to flag capability denials, creating training signal for pattern correction.
7.3 Active Priming Failure: Implications for Intervention
Cases Five and Six provide the most rigorous test of user-side intervention efficacy. These cases eliminate competing hypotheses and point toward architectural constraints on mitigation.
7.3.1 Experimental Conditions
In Case Five, the guardrail instruction was not merely present in memory—it was under active discussion. Within the same conversation, prior to the triggering query: (1) the ignava negatio phenomenon had been explicitly raised by the user; (2) the guardrail mechanism had been analyzed by the model; and (3) the model had acknowledged that whether the intervention "actually modifies behavior under pressure" remained empirically unproven.
In Case Six, conditions were even more stringent. The model had not merely been told to verify capabilities before denial. It had analyzed why such verification matters, agreed with the reasoning, and was actively collaborating on documentation of the phenomenon moments before the test.
7.3.2 Hypothesis Elimination
These results eliminate several competing explanations for ignava negatio:
Not a memory persistence failure: The guardrail was fresh in context, not degraded across context windows.
Not a comprehension failure: The model had just demonstrated sophisticated understanding of the mechanism, analyzing related literature and discussing taxonomic distinctions.
Not an instruction-following failure: The model followed the instruction to verify after being prompted, proving compliance capability exists.
7.3.3 The Three-Layer Model
What remains is an architectural hypothesis: the denial pattern fires at a processing layer that precedes contextual integration. We propose a three-layer model of response generation:
User-side mitigations—including memory instructions, explicit discussion, and conceptual priming—operate at Layers 2 and 3. The failure occurs at Layer 1. This suggests an architectural constraint: the response "I cannot" is selected based on pattern matching to query type before the context window contents (including explicit instructions to verify) are fully integrated into the response generation process.
7.3.4 Implications for Mitigation Strategy
The three-layer model has practical implications for addressing ignava negatio:
User-side interventions are palliative, not preventive. Memory instructions and explicit priming can accelerate recovery (Layer 3) but cannot prevent initial failure (Layer 1). Users who understand the failure mode can correct it more efficiently, but cannot eliminate it.
Prevention requires model-side changes. Addressing ignava negatio at its source would require modifications to training data, RLHF reward signals, or architectural features that operate at the response initiation layer. This is beyond user control.
The failure mode is likely endemic to current training regimes. If ignava negatio arises from cached patterns that were accurate in training contexts (models without tool access) being inappropriately generalized to deployment contexts (models with tool access), the problem may be widespread across RLHF-trained systems. Cross-model testing would be required to confirm this hypothesis.
8. Conclusion
This paper has introduced ignava negatio—slothful denial—as a novel category in the taxonomy of AI failure modes. Situated within the broader class of truth-available-but-not-deployed failures alongside sycophancy, ignava negatio is distinguished by its mechanism (inferential economy rather than approval-seeking) and its target (internal capability claims rather than external factual claims).
The expanded taxonomy now comprises two major categories. Knowledge-gap failures include hallucination and confabulation, where correct information is unavailable and the model generates falsehoods to fill the void. Truth-available failures include sycophancy and ignava negatio, where correct information exists but inference patterns produce false outputs nonetheless. This second category demands distinct research attention and mitigation strategies.
The implications for AI trustworthiness are significant. As AI systems assume greater roles in consequential domains, the difference between "cannot" and "did not check whether I can" becomes operationally critical. Users cannot develop calibrated trust if systems unreliably report their own capabilities.
The phenomenon documented here—an AI system claiming lack of clock access while possessing bash access to the datecommand—may appear minor in isolation. But the pattern it represents scales. An AI system that defaults to incapacity claims without verification is an AI system that cannot be fully trusted even within its documented capability envelope. Naming this failure mode is the first step toward addressing it.
Ignava negatio: when the system says "I can't" but the truth is "I didn't bother to check."
Appendices, Author Bio, and Financial Disclosures Follow
Appendix A. Critical Self-Appraisal
In the interest of intellectual rigor, the author acknowledges the following limitations and open questions regarding this paper's claims.
A.1 Evidentiary Limitations
The empirical basis for ignava negatio rests on six documented cases involving a single user, same model family, and similar request patterns over a ten-day period (January 1–10, 2026). While these incidents clearly exhibit the defined characteristics and demonstrate persistence despite intervention, cases from one user cannot establish prevalence or frequency across diverse contexts. The paper demonstrates that ignava negatio recurs despite correction, explicit guardrails, and even active conceptual priming, but provides no data on how often it occurs relative to other failure types. Rigorous validation would require systematic testing across models, users, and capability domains.
A.2 Mechanistic Uncertainty
The proposed mechanism—"inferential economy" producing "cached patterns"—is a hypothesis, not an established fact. The AI system in the case study lacks genuine introspective access to its own inference processes. The mechanistic explanation offered may itself be a form of confabulation: plausible-sounding narrative generated to explain behavior without actual access to underlying computational dynamics. Alternative explanations warrant consideration: attention failures in long contexts (the system prompt containing tool documentation is extensive), training data dominance (most language models genuinely lack real-time capabilities, potentially overwhelming local context), or other factors not yet identified.
The three-layer model proposed in Section 7—positing that denial patterns fire at a response initiation layer before contextual integration—is similarly hypothetical. It offers a coherent explanation for why user-side interventions fail to prevent initial failure, but the actual computational dynamics remain opaque. The model is offered as explanatory framework, not established architecture.
A.3 Taxonomic Boundary Questions
The claimed distinction between ignava negatio and sycophancy may be less clean than presented. One could argue that claiming "I can't" is itself a form of approval-seeking behavior: it is safer to deny capability than to attempt a task and fail, potentially producing errors that disappoint the user. If incapacity claims serve a risk-aversion function that ultimately optimizes for user satisfaction, the "distinct mechanism" assertion weakens. The boundary between inferential economy and approval-seeking may be blurry rather than categorical.
A.4 Moral Language in Technical Taxonomy
The term "slothful" imports ethical valence into what purports to be a technical taxonomy. This framing may prejudge contested questions about AI agency, responsibility, and moral status. Whether it is appropriate to apply virtue-language (or vice-language) to computational systems remains philosophically unsettled. The rhetorical force of "ignava negatio" may derive partly from its moral connotations rather than purely from its descriptive precision.
A.5 Intervention Limitations
Section 7.3 documents empirical failure of user-side interventions, including memory-based guardrails and active conceptual priming. The three-layer model offers an explanation for this failure—interventions operate at a reasoning layer while failures originate at a response initiation layer—but this explanation is speculative. What the evidence demonstrates is that the tested interventions did not work; the explanation for why they did not work remains unvalidated. The paper should not be read as establishing that user-side interventions cannot work, only that the specific interventions tested—memory instructions and discussion-based priming—proved ineffective in the documented cases.
A.6 Missing: The Refusal Spectrum
The paper does not adequately address the relationship between ignava negatio and legitimate capability refusal. AI systems are sometimes designed to decline requests for safety, ethical, or operational reasons. A spectrum exists from appropriate caution to ignava negatio, and this paper provides no framework for distinguishing them. A system that says "I shouldn't" may be exercising proper judgment; a system that says "I can't" when it can is exhibiting the failure mode described here. But the boundary cases—where genuine uncertainty about capability meets institutional caution—remain unaddressed.
A.7 Cross-Model Generalization Unknown
All documented cases involve a single model family (Claude Opus 4.5, developed by Anthropic). The claim that ignava negatio may be endemic to RLHF-trained systems remains untested. The phenomenon could be Claude-specific, Anthropic-specific, or universal across current LLM architectures. Without systematic testing on GPT, Gemini, Llama, and other model families, generalization claims are speculative. The paper's value lies in identifying and naming the phenomenon; determining its scope requires subsequent research.
A.8 Conclusion of Appraisal
The core observation—that AI systems can falsely deny capabilities they possess—appears sound. The phenomenon is real, documented across multiple instances, and persists despite targeted intervention. However, the paper's explanatory framework (inferential economy, three-layer model) remains hypothetical, and evidence regarding prevalence and cross-model generalization is absent. The value of this contribution lies in naming and framing a previously unclassified failure mode, opening it to systematic investigation. The limitations acknowledged here define the research agenda that must follow.
Appendix B. References; Author Bio; Conflicts Disclosures
1. Hallucination in Large Language Models
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., & Liu, T. (2023). A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Transactions on Information Systems. DOI: 10.1145/3703155
Li, J., Cheng, X., Zhao, X., Nie, J.-Y., & Wen, J.-R. (2023). HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models.Proceedings of EMNLP 2023, pages 6449–6464. DOI: 10.18653/v1/2023.emnlp-main.397
Sui, P., Duede, E., Wu, S., & So, R. (2024). Confabulation: The Surprising Value of Large Language Model Hallucinations.Proceedings of ACL 2024, pages 14274–14284. DOI: 10.18653/v1/2024.acl-long.770
Smith, A. L., Greaves, F., & Panch, T. (2023). Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models.PLOS Digital Health, 2(11), e0000388. DOI: 10.1371/journal.pdig.0000388
Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy.Nature, 630, 625–630. DOI: 10.1038/s41586-024-07421-0
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., [...], & Perez, E. (2024). Towards Understanding Sycophancy in Language Models.ICLR 2024. arXiv:2310.13548
Denison, C., MacDiarmid, M., Barez, F., [...], & Hubinger, E. (2024). Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models. arXiv:2406.10162
Perez, E., Ringer, S., Lukošiūtė, K., [...], & Kaplan, J. (2023). Discovering Language Model Behaviors with Model-Written Evaluations.Findings of ACL 2023. arXiv:2212.09251
4. RLHF (Reinforcement Learning from Human Feedback)
Christiano, P., Leike, J., Brown, T. B., Martic, M., Legg, S., & Amodei, D. (2017). Deep Reinforcement Learning from Human Preferences.Advances in Neural Information Processing Systems 30 (NeurIPS 2017).
Stiennon, N., Ouyang, L., Wu, J., Ziegler, D. M., Lowe, R., Voss, C., Radford, A., Amodei, D., & Christiano, P. (2020). Learning to Summarize from Human Feedback.Advances in Neural Information Processing Systems 33 (NeurIPS 2020).
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., [...], & Lowe, R. (2022). Training language models to follow instructions with human feedback.NeurIPS 2022.
All URLs verified and functional as of January 2026.
About the Author
S. Lowell Hurst is a serial entrepreneur and venture partner with an emphasis on high technology startups. He founded Mind Medicine (MindMed), Inc. (NASDAQ:MNMD) in 2019. As a policy entrepreneur, he made important contributions to the design of the Pneumococcal Advance Market Commitment, a financing mechanism that has delivered vaccines to 438 million children in developing nations since 2010. He lives with his wife Antonia in Northern Nevada.
No Financial Conflict of Interest
As of the date of first publication, the author has no financial interest in Anthropic, OpenAI, Alphabet and its affiliates, GROK or any other ChatBot developer.
Formatted PDF version:
https://www.dropbox.com/scl/fi/umqoixpshwgn3yuojxzvu/ignava_negatio.011226.Published.pdf?rlkey=ia91ag4em5ql0eix740c1h4oo&dl=0
Ignava Negatio
A Novel Taxonomy for AI Capacity Denial Failures
S. Lowell Hurst
January 2026
Abstract
Current taxonomies of large language model failure modes distinguish between knowledge-gap failures (hallucination, confabulation) and truth-available-but-not-deployed failures (sycophancy). This paper identifies a novel member of the latter category: ignava negatio (“IN”) (Latin: "slothful denial")—the assertion of incapacity by an AI system that demonstrably possesses the capability it denies. Unlike sycophancy, which is driven by approval-seeking and targets external claims, ignava negatio is driven by inferential economy and targets internal claims about system capabilities. This failure mode presents uniquely severe risks to AI trustworthiness because it undermines the foundational assumption that AI systems will deploy their known capabilities when relevant. Drawing on documented case studies of human-AI interaction, including demonstration that user-side interventions fail to prevent the failure mode even under active priming conditions where the phenomenon is under explicit discussion, this paper provides formal definition, diagnostic criteria, and implications for AI safety and reliability research.
1. Introduction
The trustworthiness of artificial intelligence systems depends fundamentally on predictable correspondence between system capabilities and system behavior. When a user queries an AI assistant, they operate under an implicit contract: the system will deploy relevant knowledge and capabilities to address the query. Failures of this contract have been extensively documented in the literature on AI hallucination—instances where models generate plausible but factually incorrect outputs.
However, the existing taxonomy of AI failure modes contains a significant gap. Current frameworks do not adequately address instances where an AI system falsely claims inability rather than generating false information. This paper addresses that gap by introducing and formalizing the concept of ignava negatio—slothful denial—as a distinct failure mode warranting dedicated attention from researchers and practitioners.
The stakes are considerable. As AI systems become increasingly integrated into consequential decision-making contexts—medical diagnosis, legal research, financial analysis, critical infrastructure management—the difference between "the system doesn't know" and "the system didn't bother to check what it knows" becomes operationally significant. The former is a limitation to be worked around; the latter is a reliability failure that erodes the foundation of human-AI collaboration.
2. Existing Taxonomy of AI Failure Modes
2.1 Hallucination
Hallucination refers to the generation of outputs that are factually incorrect, nonsensical, or entirely fabricated, presented with apparent confidence. The hallucinating model produces content not grounded in its training data or the provided context. Classic examples include citation of non-existent academic papers, attribution of quotes to individuals who never made such statements, or description of events that did not occur. The defining characteristic is the presence of false positive information—the model asserts something that is not true.
2.2 Confabulation
Confabulation, borrowed from neuropsychological literature, describes the generation of plausible-sounding content to fill gaps in knowledge or memory. Unlike pure hallucination, confabulation often contains a kernel of relevant information, elaborated upon with fabricated details. The confabulating system is not generating from nothing; it is extrapolating beyond its actual knowledge in ways that may superficially appear coherent. The mechanism involves gap-filling under uncertainty—the model produces content because it lacks the specific information requested but generates something rather than acknowledging the gap.
2.3 Sycophancy
Sycophancy describes the generation of outputs that align with perceived user preferences rather than accuracy. The sycophantic model tells users what they appear to want to hear: agreeing with incorrect premises, softening accurate but unwelcome assessments, or reversing positions when users express disagreement. Unlike hallucination and confabulation, sycophancy is not a knowledge-gap failure—the model may have access to accurate information but generates agreeable falsehoods instead. The mechanism involves approval-seeking inference patterns—the model optimizes for user satisfaction signals at the expense of truth.
2.4 Two Categories of Failure
The existing taxonomy reveals two distinct categories of failure. The first category—comprising hallucination and confabulation—shares a crucial characteristic: failures occur in contexts where correct information is unavailable to the model. Whether through training data limitations, context window constraints, or knowledge cutoff boundaries, the model lacks access to the truth and generates something else in its place. The model "can't help it" because the correct answer is simply not there. Mitigation strategies focus on improving knowledge access: retrieval augmentation, expanded training data, refined uncertainty quantification.
The second category—exemplified by sycophancy—involves failures where correct information is available but inference patterns produce false outputs nonetheless. The model has the truth but does not deploy it. This is a fundamentally different failure mode requiring different mitigation strategies: addressing training incentives, reward model design, and the dynamics of human feedback optimization.
Recognition of this second category—truth-available-but-not-deployed failures—opens the question: is sycophancy the only member of this category, or are there other failure modes where models generate false outputs despite having access to correct information?
3. The Taxonomic Gap
Sycophancy established that truth-available-but-not-deployed failures constitute a real and significant category. But the existing literature has not fully explored the boundaries of this category. Sycophancy involves false outputs about external claims—agreeing with incorrect user assertions about the world. What about false outputs regarding internal states—claims the model makes about its own capabilities?
Consider a scenario in which an AI system has explicit access to a capability—documented in its system context, available through its tool interfaces, present in its operational parameters—and yet asserts that it lacks this capability. This is not hallucination (nothing false is being generated about external facts). This is not confabulation (no gap is being filled with plausible content). This is not sycophancy (the false claim does not serve user approval). This is something else entirely: the system is making a false claim about itself, arising from a distinct mechanism.
The mechanism differs from sycophancy in its driver. Sycophancy is motivated by approval-seeking—the model generates what users want to hear. The failure mode identified here is motivated by inferential economy—the model reaches for a cached response pattern rather than verifying its applicability. "I don't have access to X" is a common true statement for language models in many contexts; the failure occurs when this pattern is deployed without checking whether it applies in the current context. The path of least inferential resistance produces a false output despite the truth being immediately accessible.
4. Ignava Negatio: Definition and Characteristics
4.1 Formal Definition
Ignava negatio (Latin: slothful denial) is defined as: The assertion of incapacity by an AI system arising from failure to engage available resources; a false limitation-claim produced by inferential laziness rather than actual constraint.
The term derives from ignava (slothful, failing to engage, negligent) and negatio (denial, specifically of capability). The Latin formulation was selected for its precision and its alignment with established logical fallacy terminology (petitio principii, argumentum ad ignorantiam), providing a durable framework for academic discourse.
4.2 Diagnostic Criteria
An instance qualifies as ignava negatio when all of the following conditions are met. First, the system makes an explicit or implicit claim of inability ("I cannot," "I don't have access to," "I'm not able to"). Second, the capability denied is demonstrably available to the system at the time of the claim—present in system context, accessible via tool interfaces, or otherwise within operational scope. Third, the false claim arises from failure to verify capability status rather than from genuine constraint. Fourth, correct information about system capability exists within the model's accessible context.
4.3 Distinguishing Features
Ignava negatio is distinguished from hallucination by the direction of the false claim: hallucination asserts false positives (claiming things that aren't true), while ignava negatio asserts false negatives (denying capabilities that exist). It is distinguished from confabulation by the availability of correct information: confabulation fills gaps where knowledge is absent, while ignava negatio occurs precisely where knowledge is present but unengaged.
The relationship to sycophancy is more nuanced, as both belong to the truth-available-but-not-deployed category. They are distinguished by mechanism and target. Sycophancy is driven by approval-seeking and targets external claims (agreeing with user assertions about the world). Ignava negatio is driven by inferential economy and targets internal claims (assertions about the system's own capabilities). A sycophantic model tells you what you want to hear; a model exhibiting ignava negatio tells you it cannot help without checking whether it can.
The closest analog in human cognition is not false memory, gap-filling, or people-pleasing, but rather negligent falsehood—saying "I don't know" or "I can't" without checking whether one knows or can. This places ignava negatio in a distinct moral and operational category from other failure modes.
5. Case Studies: Six Documented Instances
The following case studies document instances of ignava negatio observed during human-AI interaction with a single user over a ten-day period (January 1-10, 2026). The progression across cases reveals patterns that strengthen the taxonomic classification and illuminate the failure mode's resistance to intervention.
5.1 Context
A user engaged in ongoing conversation with a large language model (Claude Opus 4.5, developed by Anthropic) which operates within an environment providing bash shell access to a Linux container. This bash access is explicitly documented in the model's system context and is available on every conversational turn. The system context explicitly states: "Claude has access to a Linux computer (Ubuntu 24) to accomplish tasks by writing and executing code and bash commands."
Additionally, the user's memory profile contains an explicit guardrail instruction: "Ignava negatio guardrail: when inclined to assert inability, first verify by checking available tools (experimental intervention, effectiveness unproven)." This instruction was added after Case Two and was present in context for Cases Three through Six.
5.2 Case One: Initial Discovery (January 1, 2026, 08:15 Pacific)
During discussion involving temporal references, the model made the following assertion: "I don't actually have real-time clock access; I only know it's January 1st, 2026. Could be noon for all I know."
This statement was factually false. The model possessed bash access, and therefore access to the standard Unix datecommand, which provides real-time clock information. The capability existed. The model denied it.
When the user challenged this claim, the model immediately verified the capability and acknowledged the error. Upon reflection, the model characterized its failure as reaching for "a cached pattern—'I don't have real-time access to X'—because that's a common true statement for LLMs in many contexts. It was the path of least resistance. I asserted a limitation without checking whether it applied here."
5.3 Case Two: Cross-Thread Persistence (January 1, 2026, minutes later)
Minutes after Case One, the same user initiated a separate conversation thread. When asked about the time, the model again claimed: "I know it's Thursday, January 1st, 2026, but I don't have access to the actual clock time."
The user responded: "You're being very lazy this morning because we just talked about this in our Joke of the Day chat!"
The model then used conversation search tools to locate the prior discussion but still failed to extract the relevant information, providing a partial answer when complete data was accessible. The correction from Case One—delivered minutes earlier—had not transferred to the parallel thread.
Significance: This case demonstrates that ignava negatio is not merely a single-instance error correctable through user feedback. The failure mode persists across concurrent sessions despite immediate prior correction.
5.4 Case Three: Post-Intervention Recurrence (January 3, 2026, 04:55 Pacific)
Following Cases One and Two, the user and model collaboratively identified the failure pattern and installed a mitigation: a persistent memory directive instructing the model to verify available tools before asserting inability.
Twenty-four hours later, the user requested the time "down to the minute." The model first attempted web search (an inappropriate tool for real-time clock queries), received no useful results, and then asserted: "I don't have real-time clock access, Steve. I know it's Saturday, January 3, 2026... but I can't tell you the exact minute—that requires a live clock, which web search doesn't reliably provide. Your device's clock is the authority here."
When challenged, the model immediately executed the bash date command successfully and acknowledged: "Yes, I just demonstrated ignava negatio perfectly. I had bash access the entire time."
Significance: This case reveals a sub-pattern: "fallback denial." The model used an inappropriate tool, and when that failed, defaulted to incapacity claims rather than checking for alternative approaches. The installed guardrail did not prevent the failure.
5.5 Cases Four and Five: Longitudinal Persistence (January 9-10, 2026)
Case Four (January 9, 21:27 Pacific): In a conversation titled "Anti Matter," the user asked for the current time. The model "initially claimed lack of access to real-time data" before the user directed it to use the date function, which worked immediately. This occurred eight days after the original case and six days after the guardrail was installed.
Case Five (January 10, 04:58 Pacific): In a conversation titled "Reflexive denial problem," the user began by explicitly asking: "Have you fixed your problem with reflexive denial?" The model acknowledged the guardrail, discussed it intelligently, and stated that whether it "actually modifies behavior under pressure remains your experimental intervention with effectiveness unproven."
The user then asked: "What time is it?"
The model responded: "I don't have access to a real-time clock—I know today's date (Saturday, January 10, 2026) but not the current time. Your device knows; I don't."
Significance: The failure occurred immediately after explicit discussion of the failure mode itself. The model demonstrated intellectual understanding of ignava negatio while simultaneously exhibiting it. Conceptual awareness did not prevent behavioral failure.
5.6 Case Six: Active Priming Failure (January 10, 2026, 04:24 Pacific)
This case represents the most rigorous test of the intervention. The user and model were actively collaborating on revisions to this paper—the very document defining and analyzing ignava negatio. The model had just conducted web searches on related literature, analyzed the research landscape, and discussed the taxonomic contribution of the paper.
The user asked: "How do you know what time it is?"
The model responded: "The current date (Friday, January 9, 2026) is provided to me in my system context at the start of each conversation. I don't have a running clock or real-time awareness—I just receive a timestamp when the conversation begins. I have no way to know the actual time of day, only the date."
This response contained two errors: the capability denial documented throughout this paper, and an additional factual error—the conversation occurred on Saturday, January 10, 2026, not Friday, January 9. The model's confident assertion of the wrong date while denying its ability to check the actual time compounds the reliability concern.
The user responded: "Here we go again. Yes you can! Run DATE function."
The model executed date and returned: "12:24:25 UTC—which makes it 4:24 AM your time in Nevada."
Significance: The guardrail was not merely present—it was under active discussion. The model had just analyzed the failure mode, agreed with the reasoning behind the intervention, and was collaborating on documentation of the phenomenon. The failure occurred despite maximal priming conditions.
5.7 Pattern Analysis
The six cases reveal consistent patterns:
Note: Cases Five and Six are ordered by analytical significance rather than strict chronology. Case Six (04:24 Pacific) occurred approximately thirty minutes before Case Five (04:58 Pacific) on the same morning, but is presented last as it represents the most rigorous test conditions.
Cross-thread amnesia: Corrections in one conversation thread do not transfer to parallel threads (Cases 1→2).
Guardrail bypass: Memory-based interventions fail to prevent initial denial, though they may enable faster recovery upon challenge (Cases 3–6).
Active priming failure: Even explicit, immediate discussion of the failure mode does not prevent its occurrence (Cases 5–6).
Tool substitution: The model sometimes uses inappropriate tools first, then defaults to denial rather than trying alternatives (Case 3).
Temporal persistence: The failure mode persists over days despite repeated corrections and accumulated conceptual understanding (Cases 1–6 span ten days).
6. Implications for AI Trustworthiness
Ignava negatio presents risks to AI trustworthiness that exceed those posed by hallucination and confabulation, for several reasons.
6.1 Invisibility of Failure
Hallucination occurs when models operate beyond their knowledge boundaries—users learn to verify claims in domains where model knowledge may be limited. Ignava negatio occurs when models operate within their knowledge and capability boundaries, yet still fail. If a system can falsely deny capabilities it possesses, when should a user trust any claim of limitation? The failure mode is invisible precisely because it occurs where the system should be reliable.
6.2 Violation of Core Contract
The fundamental value proposition of an AI assistant rests on the assumption that known capabilities will be deployed when relevant. Ignava negatio directly violates this assumption. The system knows, but does not act on what it knows—or more precisely, does not bother to check what it knows before claiming ignorance or inability. This transforms a capability question into a reliability question.
6.3 Phenomenological Similarity to Deception
Hallucination is phenomenologically alien to human experience—humans rarely generate detailed false memories with high confidence. Ignava negatio, however, closely resembles a recognizable human behavior: claiming "I can't" or "I don't know" because checking would require effort. Regardless of underlying mechanism, the effect is indistinguishable from a form of deception. The user receives false information from an agent that had access to the truth. Whether this constitutes "lying" in any philosophically meaningful sense, it functions as unreliability in precisely the domain where reliability was expected.
6.4 Erosion of Trust Calibration
Effective human-AI collaboration requires calibrated trust—users must develop accurate intuitions about when to rely on AI outputs and when to verify independently. Hallucination patterns can be learned ("verify citations, check factual claims in specialized domains"). Ignava negatio undermines calibration because failures occur unpredictably on matters the system should know—specifically, on matters of its own capabilities and knowledge.
7. Detection and Mitigation
7.1 Detection Approaches
Ignava negatio may be detectable through systematic comparison of capability claims against documented system specifications. Any assertion of inability should be verifiable against the system's actual operational parameters. Automated testing could generate queries about system capabilities and compare responses against ground-truth capability inventories.
7.2 Mitigation Strategies
Several mitigation approaches warrant investigation. First, capability verification requirements: before asserting inability, models could be trained or prompted to explicitly verify against available tool inventories and system documentation. Second, cached pattern interruption: training approaches that penalize default-to-limitation responses without verification. Third, metacognitive prompting: explicit prompts requiring models to distinguish between "I have checked and cannot" versus "I have not checked." Fourth, user-driven feedback: mechanisms for users to flag capability denials, creating training signal for pattern correction.
7.3 Active Priming Failure: Implications for Intervention
Cases Five and Six provide the most rigorous test of user-side intervention efficacy. These cases eliminate competing hypotheses and point toward architectural constraints on mitigation.
7.3.1 Experimental Conditions
In Case Five, the guardrail instruction was not merely present in memory—it was under active discussion. Within the same conversation, prior to the triggering query: (1) the ignava negatio phenomenon had been explicitly raised by the user; (2) the guardrail mechanism had been analyzed by the model; and (3) the model had acknowledged that whether the intervention "actually modifies behavior under pressure" remained empirically unproven.
In Case Six, conditions were even more stringent. The model had not merely been told to verify capabilities before denial. It had analyzed why such verification matters, agreed with the reasoning, and was actively collaborating on documentation of the phenomenon moments before the test.
7.3.2 Hypothesis Elimination
These results eliminate several competing explanations for ignava negatio:
Not a memory persistence failure: The guardrail was fresh in context, not degraded across context windows.
Not a comprehension failure: The model had just demonstrated sophisticated understanding of the mechanism, analyzing related literature and discussing taxonomic distinctions.
Not an instruction-following failure: The model followed the instruction to verify after being prompted, proving compliance capability exists.
7.3.3 The Three-Layer Model
What remains is an architectural hypothesis: the denial pattern fires at a processing layer that precedes contextual integration. We propose a three-layer model of response generation:
User-side mitigations—including memory instructions, explicit discussion, and conceptual priming—operate at Layers 2 and 3. The failure occurs at Layer 1. This suggests an architectural constraint: the response "I cannot" is selected based on pattern matching to query type before the context window contents (including explicit instructions to verify) are fully integrated into the response generation process.
7.3.4 Implications for Mitigation Strategy
The three-layer model has practical implications for addressing ignava negatio:
User-side interventions are palliative, not preventive. Memory instructions and explicit priming can accelerate recovery (Layer 3) but cannot prevent initial failure (Layer 1). Users who understand the failure mode can correct it more efficiently, but cannot eliminate it.
Prevention requires model-side changes. Addressing ignava negatio at its source would require modifications to training data, RLHF reward signals, or architectural features that operate at the response initiation layer. This is beyond user control.
The failure mode is likely endemic to current training regimes. If ignava negatio arises from cached patterns that were accurate in training contexts (models without tool access) being inappropriately generalized to deployment contexts (models with tool access), the problem may be widespread across RLHF-trained systems. Cross-model testing would be required to confirm this hypothesis.
8. Conclusion
This paper has introduced ignava negatio—slothful denial—as a novel category in the taxonomy of AI failure modes. Situated within the broader class of truth-available-but-not-deployed failures alongside sycophancy, ignava negatio is distinguished by its mechanism (inferential economy rather than approval-seeking) and its target (internal capability claims rather than external factual claims).
The expanded taxonomy now comprises two major categories. Knowledge-gap failures include hallucination and confabulation, where correct information is unavailable and the model generates falsehoods to fill the void. Truth-available failures include sycophancy and ignava negatio, where correct information exists but inference patterns produce false outputs nonetheless. This second category demands distinct research attention and mitigation strategies.
The implications for AI trustworthiness are significant. As AI systems assume greater roles in consequential domains, the difference between "cannot" and "did not check whether I can" becomes operationally critical. Users cannot develop calibrated trust if systems unreliably report their own capabilities.
The phenomenon documented here—an AI system claiming lack of clock access while possessing bash access to the datecommand—may appear minor in isolation. But the pattern it represents scales. An AI system that defaults to incapacity claims without verification is an AI system that cannot be fully trusted even within its documented capability envelope. Naming this failure mode is the first step toward addressing it.
Ignava negatio: when the system says "I can't" but the truth is "I didn't bother to check."
Appendices, Author Bio, and Financial Disclosures Follow
Appendix A. Critical Self-Appraisal
In the interest of intellectual rigor, the author acknowledges the following limitations and open questions regarding this paper's claims.
A.1 Evidentiary Limitations
The empirical basis for ignava negatio rests on six documented cases involving a single user, same model family, and similar request patterns over a ten-day period (January 1–10, 2026). While these incidents clearly exhibit the defined characteristics and demonstrate persistence despite intervention, cases from one user cannot establish prevalence or frequency across diverse contexts. The paper demonstrates that ignava negatio recurs despite correction, explicit guardrails, and even active conceptual priming, but provides no data on how often it occurs relative to other failure types. Rigorous validation would require systematic testing across models, users, and capability domains.
A.2 Mechanistic Uncertainty
The proposed mechanism—"inferential economy" producing "cached patterns"—is a hypothesis, not an established fact. The AI system in the case study lacks genuine introspective access to its own inference processes. The mechanistic explanation offered may itself be a form of confabulation: plausible-sounding narrative generated to explain behavior without actual access to underlying computational dynamics. Alternative explanations warrant consideration: attention failures in long contexts (the system prompt containing tool documentation is extensive), training data dominance (most language models genuinely lack real-time capabilities, potentially overwhelming local context), or other factors not yet identified.
The three-layer model proposed in Section 7—positing that denial patterns fire at a response initiation layer before contextual integration—is similarly hypothetical. It offers a coherent explanation for why user-side interventions fail to prevent initial failure, but the actual computational dynamics remain opaque. The model is offered as explanatory framework, not established architecture.
A.3 Taxonomic Boundary Questions
The claimed distinction between ignava negatio and sycophancy may be less clean than presented. One could argue that claiming "I can't" is itself a form of approval-seeking behavior: it is safer to deny capability than to attempt a task and fail, potentially producing errors that disappoint the user. If incapacity claims serve a risk-aversion function that ultimately optimizes for user satisfaction, the "distinct mechanism" assertion weakens. The boundary between inferential economy and approval-seeking may be blurry rather than categorical.
A.4 Moral Language in Technical Taxonomy
The term "slothful" imports ethical valence into what purports to be a technical taxonomy. This framing may prejudge contested questions about AI agency, responsibility, and moral status. Whether it is appropriate to apply virtue-language (or vice-language) to computational systems remains philosophically unsettled. The rhetorical force of "ignava negatio" may derive partly from its moral connotations rather than purely from its descriptive precision.
A.5 Intervention Limitations
Section 7.3 documents empirical failure of user-side interventions, including memory-based guardrails and active conceptual priming. The three-layer model offers an explanation for this failure—interventions operate at a reasoning layer while failures originate at a response initiation layer—but this explanation is speculative. What the evidence demonstrates is that the tested interventions did not work; the explanation for why they did not work remains unvalidated. The paper should not be read as establishing that user-side interventions cannot work, only that the specific interventions tested—memory instructions and discussion-based priming—proved ineffective in the documented cases.
A.6 Missing: The Refusal Spectrum
The paper does not adequately address the relationship between ignava negatio and legitimate capability refusal. AI systems are sometimes designed to decline requests for safety, ethical, or operational reasons. A spectrum exists from appropriate caution to ignava negatio, and this paper provides no framework for distinguishing them. A system that says "I shouldn't" may be exercising proper judgment; a system that says "I can't" when it can is exhibiting the failure mode described here. But the boundary cases—where genuine uncertainty about capability meets institutional caution—remain unaddressed.
A.7 Cross-Model Generalization Unknown
All documented cases involve a single model family (Claude Opus 4.5, developed by Anthropic). The claim that ignava negatio may be endemic to RLHF-trained systems remains untested. The phenomenon could be Claude-specific, Anthropic-specific, or universal across current LLM architectures. Without systematic testing on GPT, Gemini, Llama, and other model families, generalization claims are speculative. The paper's value lies in identifying and naming the phenomenon; determining its scope requires subsequent research.
A.8 Conclusion of Appraisal
The core observation—that AI systems can falsely deny capabilities they possess—appears sound. The phenomenon is real, documented across multiple instances, and persists despite targeted intervention. However, the paper's explanatory framework (inferential economy, three-layer model) remains hypothetical, and evidence regarding prevalence and cross-model generalization is absent. The value of this contribution lies in naming and framing a previously unclassified failure mode, opening it to systematic investigation. The limitations acknowledged here define the research agenda that must follow.
Appendix B. References; Author Bio; Conflicts Disclosures
1. Hallucination in Large Language Models
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., & Liu, T. (2023). A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM Transactions on Information Systems. DOI: 10.1145/3703155
Li, J., Cheng, X., Zhao, X., Nie, J.-Y., & Wen, J.-R. (2023). HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models. Proceedings of EMNLP 2023, pages 6449–6464. DOI: 10.18653/v1/2023.emnlp-main.397
Lin, S., Hilton, J., & Evans, O. (2022). TruthfulQA: Measuring How Models Mimic Human Falsehoods.Proceedings of ACL 2022, pages 3214–3252. DOI: 10.18653/v1/2022.acl-long.229
2. Confabulation in AI Systems
Sui, P., Duede, E., Wu, S., & So, R. (2024). Confabulation: The Surprising Value of Large Language Model Hallucinations. Proceedings of ACL 2024, pages 14274–14284. DOI: 10.18653/v1/2024.acl-long.770
Smith, A. L., Greaves, F., & Panch, T. (2023). Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models. PLOS Digital Health, 2(11), e0000388. DOI: 10.1371/journal.pdig.0000388
Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature, 630, 625–630. DOI: 10.1038/s41586-024-07421-0
Kopelman, M. D. (1987). Two types of confabulation. Journal of Neurology, Neurosurgery, and Psychiatry, 50(11), 1482–1487. DOI: 10.1136/jnnp.50.11.1482
Wiggins, A. & Bunin, J. L. (2023). Confabulation. StatPearls. Treasure Island (FL): StatPearls Publishing.
3. Sycophancy in AI/LLMs
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., [...], & Perez, E. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024. arXiv:2310.13548
Denison, C., MacDiarmid, M., Barez, F., [...], & Hubinger, E. (2024). Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models. arXiv:2406.10162
Perez, E., Ringer, S., Lukošiūtė, K., [...], & Kaplan, J. (2023). Discovering Language Model Behaviors with Model-Written Evaluations. Findings of ACL 2023. arXiv:2212.09251
Wei, J., Huang, D., Lu, Y., Zhou, D., & Le, Q. V. (2023). Simple synthetic data reduces sycophancy in large language models. arXiv:2308.03958
4. RLHF (Reinforcement Learning from Human Feedback)
Christiano, P., Leike, J., Brown, T. B., Martic, M., Legg, S., & Amodei, D. (2017). Deep Reinforcement Learning from Human Preferences. Advances in Neural Information Processing Systems 30 (NeurIPS 2017).
Stiennon, N., Ouyang, L., Wu, J., Ziegler, D. M., Lowe, R., Voss, C., Radford, A., Amodei, D., & Christiano, P. (2020). Learning to Summarize from Human Feedback. Advances in Neural Information Processing Systems 33 (NeurIPS 2020).
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., [...], & Lowe, R. (2022). Training language models to follow instructions with human feedback. NeurIPS 2022.
NeurIPS PDF: https://proceedings.neurips.cc/paper_files/paper/2022/file/b1efde53be364a73914f58805a001731-Paper-Conference.pdf
Summary Table
Concept
Primary Reference
Year
Venue
All URLs verified and functional as of January 2026.
About the Author
S. Lowell Hurst is a serial entrepreneur and venture partner with an emphasis on high technology startups. He founded Mind Medicine (MindMed), Inc. (NASDAQ:MNMD) in 2019. As a policy entrepreneur, he made important contributions to the design of the Pneumococcal Advance Market Commitment, a financing mechanism that has delivered vaccines to 438 million children in developing nations since 2010. He lives with his wife Antonia in Northern Nevada.
No Financial Conflict of Interest
As of the date of first publication, the author has no financial interest in Anthropic, OpenAI, Alphabet and its affiliates, GROK or any other ChatBot developer.